v1.3 — Usage analytics dashboard
Per-model, per-key, per-team spend charts on a single page. Filter by date range, export CSV, drill into outliers. Numbers render in Geist Mono — easy on the eyes during the monthly close.
Knowing what you spent matters as much as making the call. v1.3 ships the usage analytics page — one screen, three dimensions, all the way down to the individual request.
Three slices, one page
At /{org}/usage you get:
- Per-model — total spend, request count, average latency, p95 latency. Sorted by spend descending so the budget drivers surface first.
- Per-virtual-key — same metrics by key. The most-used key gets a hit-rate badge; the most expensive key gets a cost-share percentage. Pair with budgets at
/{org}/budgetsfor hard caps. - Per-team — for orgs with multiple teams (Rule #26), spend rolls up by team. Useful for chargeback or budget reviews.
Time filters + export
Default range is the trailing 7 days, with quick toggles for 24h / 7d / 30d / 90d / billing-cycle. Custom ranges via the date picker. Every view exports to CSV with a single click — no JSON munging, no awk one-liners.
Where the numbers come from
LiteLLM owns cost (Rule #4). Every assistant response carries an x-litellm-response-cost header which we record into LiteLLM_SpendLogs and roll into the usage charts. We never compute cost ourselves — if the provider returns $0, we record $0. The dashboard reflects exactly what was billed against your credit balance, no rounding, no extrapolation.
Numbers render in Geist Mono so columns line up cleanly. Charts are Recharts wrapped in <NemoChart>. All amounts are pre-platform-fee — the platform fee was already added at credit purchase, never deducted from your balance per-request.